DeepMind's AlphaProof Code formally verifies key Signal protocol guarantees

The four-week project found no new flaws in the parts examined and caught old, fixed bugs. Google declined to share the cost, and whether it can scale at a workable cost is an open question, Wired says.

SecurityPublished Updated
Image published by Google DeepMind with “Introducing AlphaProof Code”
Image: Google DeepMind · Source

Correction, . An earlier version said Google would not say whether the approach scales. Wired reports that Google declined to share the cost and to comment on commercial plans; whether it can scale at a workable cost is an open question, not something Google refused to answer.

What happened
Google DeepMind used an agentic AI framework, AlphaProof Code, to mathematically prove several core cryptographic guarantees of the Signal protocol.
What’s new
AI agents carried out much of the proof work in four weeks, extending earlier, narrower human-led verification of Signal's code.
Why it matters
The Signal protocol also secures WhatsApp and Google Messages, so independent proof of its guarantees bears on a large share of encrypted messaging.

Google DeepMind researchers have used agentic AI to formally verify a number of the Signal protocol's key cryptographic guarantees, Wired reports. Formal verification means using mathematics to prove that software behaves as specified, rather than testing it for bugs. The work relied on a new DeepMind framework called AlphaProof Code and took four weeks as a proof-of-concept project.

The Signal protocol is used by the Signal app, WhatsApp and Google Messages. A result about the protocol therefore reaches well beyond a single messaging service.

What the proof covers, and what it does not

Both outlets' headlines say the analysis proved Signal's protocol secure (9to5Mac adds WhatsApp), which is broader than their own text: both say it covered key guarantees of the protocol, not the Signal app as a whole. It describes formal verification as guaranteeing that there is no situation in which a protocol will fail.

Wired's account is narrower. It says DeepMind verified a number of the protocol's key guarantees and that the analysis was limited in scope. It did not, for example, cover the whole Signal mobile app. 9to5Mac also notes that the validation applies to the protocol, not the app.

Within those limits, the analysis found no new vulnerabilities in the portions of the protocol it examined. As a check, the researchers ran AlphaProof Code against older Signal versions with known, long-fixed vulnerabilities. The tool flagged those versions as failing verification. DeepMind has published the specification and proof (google-deepmind.github.io/code-verification). Its paper says the proofs connect protocol-level security guarantees to libsignal, Signal's reference implementation, and that it had to repair errors in published analyses of the Double Ratchet and related protocol models before the proofs would go through.

Human review and a head start

The speed came partly from earlier work. According to Wired, DeepMind built on past formal verification carried out by the Signal Foundation without AI. Signal CTO Ehren Kret said in a statement that the earlier efforts covered smaller portions of the code, while DeepMind's approach allows significant expansion.

DeepMind research scientist Swarat Chaudhuri described the method as a human-AI collaboration process. In it, agents contribute and self-correct at most stages but seek human input and review at key points. He contrasted it with a setup in which AI writes both the goals and the proofs, telling Wired, "There’s a middle ground."

Open questions on cost and scale

Both outlets say the practicality of the approach is unresolved, though 9to5Mac's account relies on Wired's reporting. Wired reports that DeepMind did the analysis free of charge as a research project. Google declined to disclose the cost or to say whether it plans a commercial formal verification product. 9to5Mac notes that the resources required have not been disclosed and describes the method as extremely computationally demanding.

Chaudhuri told Wired there is a long distance to go before the technique can scale to all the world's software. DeepMind vice president of research Pushmeet Kohli pointed to banks and other companies that want provable assurances about their code as potential beneficiaries. Wired says whether AI-automated verification can scale at a workable cost remains an open question across the industry.